I have a timeseries with enough data points that it's difficult to store it all in memory at once, but would like to solve a linear system of equations using all of it, so I'm looking for a way to solve an initial LU factorization, and then read in pieces at a time from the file to refine the solution to something more accurate.
Do any of you know of a specific technique for doing this sort of thing? I've looked into iterative refinement (e.g.) but that seems to be more for improving accuracy with the same data rather than new data.
Alternatively, would it be best to just break the $m$ by $n$data points into $n/m$ square matrices and average the resulting coefficients?
Thanks
